How Anthropic Used AI to Double Its Own Development Speed

According to an independent evaluation by the nonprofit organization Model Evaluation and Threat Research (METR), cited by Olhar Digital, Anthropic's model capabilities have advanced at a 1.9x faster rate since October 2025. The core driver isn't just raw compute power, but researchers actively deploying AI systems to handle tasks like writing code and executing experiments.

What does a 1.9x acceleration mean for tech timelines?

In practical engineering terms, a nearly doubled development velocity cuts experimental cycles almost in half. Tasks that previously required days of manual scripting and trial-and-error by human researchers are now offloaded to internal AI assistants. While METR notes that some informal internal estimates claimed speedups of up to 12x, these subjective claims lack rigorous backing, keeping the validated baseline at a steady 1.9x multiplier.

How is AI changing daily research workflows?

The study emphasizes that current gains stem from AI acting as an advanced assistant rather than operating autonomously. Researchers use models to draft code segments, set up test environments, and analyze performance data. This eliminates routine bottlenecks, allowing engineers to focus on architectural decisions rather than boilerplate coding.

• AI writes baseline code for new model experiments.
• Researchers run automated validation tests significantly faster.
• Engineers spend more time on safety frameworks rather than syntax.

Is recursive self-improvement finally here?

This dynamic touches directly on Recursive Self-Improvement (RSI), where systems aid in building their successors. METR found no evidence of fully autonomous AI breakthroughs driving the training pipeline entirely on their own yet. However, the feedback loop between human researchers and faster AI tools creates a cumulative momentum that makes tracking safety milestones increasingly urgent.

The Verdict: What this acceleration means for the industry

For professionals relying on AI tools, this 1.9x velocity bump signals that update cycles for models like Claude will likely compress. If the trend holds, waiting six months for a major capability jump could shrink into a much tighter window. Yet, as CEO Dario Amodei has previously argued, faster development cycles also heighten the challenge of keeping guardrails effective as capability curves steepen.